// RUN: tf-quant-opt %s -split-input-file -quant-lift-quantizable-spots-as-functions -quant-prepare-quantize-drq -quant-quantize='weight-quantization=true' -verify-each=false | FileCheck %s

// -----

module {
  func.func @matmul(%arg0: tensor<1x2x2x3xf32>) -> (tensor<*xf32>) {
    %cst_0 = "tf.Const"() {value = dense<0.000000e+00> : tensor<2x1024xf32>} : () -> tensor<2x1024xf32>

  %0 = corert.get_op_handler %arg0 "cpu"
  // CHECK: %[[RESULT:.*]] = corert.executeop(%[[ARG_0:.*]]) "tf.MatMul"(%[[ARG_1:.*]], %[[ARG_1]]) {T = f32, transpose_a = false, transpose_b = false} : 1
  %1 = corert.executeop(%0) "tf.MatMul"(%arg1, %arg1) {T = f32, device = "cpu", transpose_a = false, transpose_b = false} : 1
  tfrt.return %arg0, %1 : !tfrt.chain, !corert.tensorhandle

# RUN: tf_tfl_translate -tf-input-arrays=Placeholder,Placeholder_1 -tf-input-shapes=2,5,3:3,7 -tf-input-data-types=DT_FLOAT,DT_FLOAT -tf-output-arrays=MatMul -unfold_batchmatmul=true -output-mlir %s -o - 2>&1 | FileCheck %s

node {
  name: "Placeholder"
  op: "Placeholder"
  attr {
    key: "dtype"
    value {
      type: DT_FLOAT
    }
  }
  attr {
    key: "shape"
    value {
      shape {
        dim {
          size: 2
        }

// CHECK: %[[IDENTITY:.*]] = "tf.Identity"(%cst) : (tensor<2x1024xi8>) -> tensor<2x1024xi8>
// CHECK: %[[MATMUL:.*]] = "tf.XlaDotV2"(%arg0, %[[IDENTITY]]) <{dimension_numbers = "\12\01\00\0A\01\03", precision_config = ""}> {device = ""} : (tensor<1x2x2x2xbf16>, tensor<2x1024xi8>) -> tensor<1x2x2x1024xbf16>

  // CHECK: tfrt_fallback_async.executeop key(2) cost({{.*}}) device("/device:CPU:0") "tf.MatMul"
  %0 = "tf.ReadVariableOp"(%arg1) {device = "/device:CPU:0", dtype = f32} : (tensor<!tf_type.resource<tensor<1x3xf32>>>) -> tensor<1x3xf32>
  %1 = "tf.MatMul"(%arg0, %0) {T = f32, device = "/device:CPU:0", transpose_a = false, transpose_b = false} : (tensor<3x1xf32>, tensor<1x3xf32>) -> tensor<3x3xf32>

  %cst = "tf.Const"() {value = dense<0.000000e+00> : tensor<512x512xf32>} : () -> tensor<512x512xf32>
  %out_1 = "tf.MatMul"(%arg0, %cst) {
    device = "", transpose_a = false, transpose_b = false
  } : (tensor<1x12x12x512xf32>, tensor<512x512xf32>) -> tensor<*xf32>
  %out_2 = "tf.MatMul"(%arg0, %arg0) {
    device = "", transpose_a = false, transpose_b = true
  } : (tensor<1x12x12x512xf32>, tensor<1x12x12x512xf32>) -> tensor<*xf32>

  ASSERT_EQ(TF_OK, TF_GetCode(status)) << TF_Message(status);

  TFE_Op* matmul = MatMulOp(ctx, h0_task1, h1_task1);
  TFE_OpSetDevice(matmul, remote_device_name, status);
  EXPECT_EQ(TF_OK, TF_GetCode(status)) << TF_Message(status);

  TFE_TensorHandle* retvals[1];
  int num_retvals = 1;
  TFE_Execute(matmul, &retvals[0], &num_retvals, status);
  EXPECT_EQ(TF_OK, TF_GetCode(status)) << TF_Message(status);

category names correspond to generated source file names, and should be
consistent with the original source files registering each operator. For example
since `REGISTER_OP("MatMul")` appears in ***core/math_ops.cc***, the "MatMul"
operator in the script should be in the "math" category, and it will be
generated in the output file `c/experimental/ops/math_ops.cc`.

// CHECK-NEXT:    [[TENSOR:%.*]] = "tf_mlrt.tf_await"([[FURTURE]]) : (!mlrt.future) -> tensor<3x1xf32>
// CHECK-NEXT:    "tf.MatMul"(%arg0, [[TENSOR]]) : (tensor<1x3xf32>, tensor<3x1xf32>) -> tensor<1x1xf32>

# CHECK-NEXT: [[R1:%.*]] = "tf.ReadVariableOp"([[ARG1]]) {{{.*}}} : (tensor<!tf_type.resource<tensor<3x5xf32>>>) -> tensor<3x5xf32>
# CHECK-NEXT: [[R2:%.*]] = "tf.MatMul"([[R0]], [[R1]]) <{{{.*}}}> {{{.*}}} : (tensor<5x3xf32>, tensor<3x5xf32>) -> tensor<5x5xf32>


def Test():

  x = tf.compat.v1.get_variable(
      name='x',
      shape=(5, 3),